Google’s Gary Illyes answered questions during the recent Search Central Live Deep Dive in Asia about whether or not they use the new Multi‑Vector Retrieval via Fixed‑Dimensional Encodings (MUVERA) retrieval method and also if they’re using Graph Foundation Models.
MUVERA
Google recently announced MUVERA in a blog post and a research paper: a method that improves retrieval by turning complex multi-vector search into fast single-vector search. It compresses sets of token embeddings into fixed-dimensional vectors that closely approximate their original similarity. This lets it use optimized single-vector search methods to quickly find good candidates, then re-rank them using exact multi-vector similarity. Compared to older systems like PLAID, MUVERA is faster, retrieves fewer candidates, and still improves recall, making it a practical solution for large-scale retrieval.
The key points about MUVERA are:
- MUVERA converts multi-vector sets into fixed vectors using Fixed Dimensional Encodings (FDEs), which are single-vector representations of multi-vector sets.
- These FDEs (Fixed Dimensional Encodings) match the original multi-vector comparisons closely enough to support accurate retrieval.
- MUVERA retrieval uses MIPS (Maximum Inner Product Search), an established search technique used in retrieval, making it easier to deploy at scale.
- Reranking: After using fast single-vector search (MIPS) to quickly narrow down the most likely matches, MUVERA re-ranks them using Chamfer similarity, a more detailed multi-vector comparison method. This final step restores the full accuracy of multi-vector retrieval, so you get both speed and precision.
- MUVERA is able to find more of the precisely relevant documents with a lower processing time than the state-of-the-art retrieval baseline (PLAID) it was compared to.
Google Confirms That They Use MUVERA
José Manuel Morgal (LinkedIn profile) related his question to Google’s Gary Illyes and his response was to jokingly ask what MUVERA was and then he confirmed that they use a version of it:
This is how the question and answer was described by José:
“An article has been published in Google Research about MUVERA and there is an associated paper. Is it currently in production in Search?
His response was to ask me what MUVERA was haha and then he commented that they use something similar to MUVERA but they don’t name it like that.”
Does Google Use Graph Foundation Models (GFMs)?
Google recently published a blog announcement about an AI breakthrough called a Graph Foundation Model.
Google’s Graph Foundation Model (GFM) is a type of AI that learns from relational databases by turning them into graphs, where rows become nodes and the connections between tables become edges.
Unlike older models (machine learning models and graph neural networks (GNNs)) that only work on one dataset, GFMs can handle new databases with different structures and features without retraining on the new data. GFMs use a large AI model to learn how data points relate across tables. This lets GFMs find patterns that regular models miss, and they perform much better in tasks like detecting spam in Google’s scaled systems. GFMs are a big step forward because they bring foundation-model flexibility to complex structured data.
Graph Foundation Models represent a notable achievement because their improvements are not incremental. They are an order-of-magnitude improvement, with performance gains of 3x to 40x in average precision.
José next asked Illyes if Google uses Graph Foundation Models and Gary again jokingly feigned not knowing what José was talking about.
He related the question and answer:
“An article has been published in Google Research about Graph Foundation Models for data, this time there are not paper associated with it. Is it currently in production in Search?
His answer was the same as before, asking me what Graph Foundation Models for data was, and he thought it was not in production. He did not know because there are not associated paper and on the other hand, he commented me that he did not control what is published in Google Research blog.”
Gary expressed his opinion that Graph Foundation Model was not currently used in Search. At this point, that’s the best information we have.
See also: Google’s New Graph Foundation Model Improves Precision By Up To 40X
Is GFM Ready For Scaled Deployment?
The official Graph Foundation Model announcement says it was tested in an internal task, spam detection in ads, which strongly suggests that real internal systems and data were used, not just academic benchmarks or simulations.
Here is what Google’s announcement relates:
“Operating at Google scale means processing graphs of billions of nodes and edges where our JAX environment and scalable TPU infrastructure particularly shines. Such data volumes are amenable for training generalist models, so we probed our GFM on several internal classification tasks like spam detection in ads, which involves dozens of large and connected relational tables. Typical tabular baselines, albeit scalable, do not consider connections between rows of different tables, and therefore miss context that might be useful for accurate predictions. Our experiments vividly demonstrate that gap.”
Takeaways
Google’s Gary Illyes confirmed that a form of MUVERA is in use at Google. His answer about GFM seemed to be expressed as an opinion, so it’s somewhat less clear, as it’s related as Gary saying that he thinks it’s not in production.
Featured Image by Shutterstock/Krakenimages.com
